Anthropic Launches Claude Sonnet 5: The Most Agentic Sonnet Yet

Anthropic released Claude Sonnet 5 on June 30, positioning it as the most agentic Sonnet model ever built. The model can plan, use tools like browsers and terminals, browse the web, write code, and run autonomously at a level that previously required much larger and more expensive models. On benchmarks like BrowseComp (agentic search) and OSWorld-Verified (computer use), Sonnet 5 closely approaches Opus 4.8 while offering dramatically better cost efficiency.

The model is available immediately across all plans — Free, Pro, Max, Team, and Enterprise — and is also accessible in Claude Code and via the Claude Platform API. Anthropic introduced introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026, after which standard pricing settles at $3 and $15 per million tokens respectively. Early testers reported that Sonnet 5 completes complex multi-step tasks where previous Sonnet models would stop short, and frequently checks its own output without explicit prompting.

On the safety front, Sonnet 5 shows lower rates of hallucination and sycophancy than Sonnet 4.6, and better resists prompt injection attacks. Notably, the model was never able to develop a full working software exploit during testing, scoring 0% on that metric — significantly safer than Opus 4.8 and Mythos 5. Anthropic has enabled real-time cyber safeguards on Sonnet 5 by default.

Why it matters: Sonnet 5 makes agentic AI capabilities accessible at a price point that changes the economics for developers and enterprises, potentially accelerating the shift from AI chatbots to autonomous AI agents in production workflows.